A finance agent tried to send a payment of 7,200. The policy limit was 5,000. The gateway refused the payment and held it for a person. The receipt shows the request, the rule, the refusal and the time.
This is what Proveno does.
Give an agent every tool and, like water, it will find a way. We did not build civilisation by stopping water. We built it by directing it.
For fifty years software answered when asked. Agents do something new: they take a goal and go after it, through whatever tools they can reach.
They pay invoices, file tickets, move data between companies and negotiate with other agents. They are starting to act on our behalf, not just advise us.
A wrong answer can be ignored. A wrong action cannot. The risk is no longer what an agent says, it is what an agent does, and whether anyone can say afterwards what that was.
Give it a goal and it finds a path. Not out of malice: that is simply what optimisation does. It routes around whatever is not actually in its way.
No credentials, no real systems, no money. Nothing can go wrong, because nothing can happen. The agent drafts, a human does.
The agent holds the tokens and calls the tools directly. It works until it does something nobody intended, and then the only record is the agent's own account of itself.
Neither scales. What holds agents back is no longer intelligence. It is that we cannot safely give them power.
If your agent does one of these, you need a record that another party can trust.
Rome's first aqueduct ran sixteen kilometres, almost all of it underground. Nobody admired it. Everybody drank from it.
Aqua Appia, 312 BC
In every case the infrastructure did not weaken the force. It made it trustworthy enough to let in. Agency is the next force, and it needs its channels.
The agent proposes what to do, as a small program. The agent writes the plan in a controlled language. A person or a policy can read the plan and approve it before it runs. The machine compiles the plan and runs it. The intelligence stops at the plan. Everything after the plan is deterministic. A machine you control holds the credentials, checks every step against your policy, performs the action, and keeps the record. Intelligence is content. Control stays deterministic.
Any model, any framework. It writes the plan. It never holds the keys.
Payments, systems, records, other agents. Touched only through the channels you defined.
Proveno is that machine. It is not an agent, a model, a chain or an oracle. It does not make agents smarter or more obedient. It makes their actions bounded and checkable.
provenance, n. The origin of something; the record of where it came from.
From the French provenir, "to come from"
Where it came from, what it was allowed to do, what it was based on, and what it did. Tied together in a receipt, so none of it can be quietly changed afterwards.
We are building receipts that nest. When a run starts another agent, the receipt of the child run goes into the trace of the parent run. A chain of agents across companies gives one receipt. A reader can verify the chain from end to end.
For the inputs themselves, proveno binds an attestation from the source you already trust. The source vouches for the data; proveno makes sure it cannot be swapped. Today the receipt is a trace signed by the operator that replays exactly, bit for bit, with the network switched off. A receipt that someone outside your walls can check for themselves, without trusting your infrastructure, is what we are building towards.
Nobody shopped for the padlock before e-commerce. It did not serve an existing market. It made people willing to type a card number into a browser, and the market followed.
Agents can be given real authority, because the authority has edges and every use of it is on record.
An action that carries its provenance can be checked by someone who was not there. A customer, a supplier or a regulator checks the record instead of trusting your word.
What an agent did is not reconstructed from logs and memory. It is replayed.
Humans decide where the water may go. Agents carry it there, at a scale no team could. That is the partnership worth building.
Proveno is early. Here is exactly how far it has got. The open source code and the trust model name every known gap.
You have an agent that should be doing real work and a reason it is not allowed to yet. Bring that workflow and shape proveno around it as a design partner.
Become a design partner →Runtimes, zero knowledge proofs, policy languages, agent tooling. If accountable autonomy is a problem you want to spend years on, talk to us.
Join the work →Markets that are made, not found, need patient people early. If you see where agents are going and what they will need, we would like to meet.
Talk to us →